Token pricing is the cheapest and most predictable part of running an AI agent. At current Anthropic API rates, a support agent handling 30,000 conversation turns a month costs roughly $1,350 in tokens — and about $378 once prompt caching is switched on. The bill that actually surprises people is everything the invoice does not itemise.
This is the arithmetic nobody publishes, worked through with real numbers and stated assumptions.
What does a single agent turn actually cost?
Multiply your input and output tokens by the per-million rate and you have the floor. Published rates differ by roughly five times across a single vendor's range, so model choice moves the number more than almost any prompt change you could make.
| Model | Input / 1M | Output / 1M |
|---|---|---|
| Claude Opus 5 | $5.00 | $25.00 |
| Claude Sonnet 5 | $2.00 | $10.00 |
| Claude Haiku 4.5 | $1.00 | $5.00 |
Rates above are Anthropic's first-party API pricing as published. Other vendors price differently, and partner platforms like Bedrock and Vertex bill on their own schedules — check before budgeting against any of them.
A worked example. Assume a retrieval-grounded support agent: a system prompt, tool definitions and retrieved documents totalling 20,000 input tokens per turn, producing 500 output tokens. On Sonnet 5 that is $0.04 in and $0.005 out — about 4.5 cents a turn.
At 10,000 conversations a month averaging three turns each, that is 30,000 turns and roughly $1,350 a month. Every figure here is an assumption, stated so you can substitute your own; none of it is drawn from a specific engagement.
Why is the real bill bigger than the token maths?
Because the token maths assumes one clean pass, and production never gives you one. Four multipliers sit between the calculation above and the invoice, and they compound rather than add.
Context grows as the conversation does. You resend the full history on every turn, so turn five costs multiples of turn one. A ten-turn conversation is not ten times a single turn — it is closer to the sum of an arithmetic series.
Retries and failures bill normally. A malformed tool call, a timeout, a response that fails validation: each one costs full price and produces nothing. Budget a realistic failure rate rather than assuming the happy path.
Agentic loops multiply turns invisibly. An agent that calls three tools before answering has made four model calls, not one. The user sees a single reply; you pay for the whole loop — which is why we scope agent engineering work by loop depth rather than by feature count.
Evals cost money too. Running a golden dataset on every release is the right practice and a real line item. So is the shadow-mode period where the agent runs alongside humans without acting.
What costs never appear on the API bill?
The ones that decide whether the agent survives. Roughly 88% of AI agents never reach production, and the attributed causes are infrastructure gaps (41%), governance and security barriers (38%) and ROI measurement failures (33%) — not model spend.
Human review. Most defensible automations keep a person on the exceptions. If 10% of 30,000 turns escalate, that is 3,000 reviews a month, and reviewer time is usually the largest single cost in the whole system.
The data layer underneath. An agent reading from systems that disagree with each other produces confident wrong answers. Fixing that is a business systems and data foundations problem, and it is frequently the majority of a first build.
Observability and evals. Tracing, dashboards, golden datasets and the engineering time to maintain them. Skipping this is how teams end up unable to tell whether a release made things worse.
Maintenance. Prompts drift as the product changes, upstream schemas move, and model versions are deprecated. An unmaintained agent degrades quietly, which is worse than failing loudly.
Governance, where it applies. Systems touching hiring, credit, pricing or clinical care carry oversight and evidence obligations. Designing those in costs far less than retrofitting them — which is what AI governance and assurance exists to handle.
How do you cut the bill without cutting quality?
Take the free wins before the tradeoffs. Two mechanisms cut token spend substantially without touching output quality at all, and most teams have enabled neither.
Prompt caching is the largest single lever. Cached prefix tokens are read at roughly a tenth of normal cost, with cache writes at about 1.25×. Return to the worked example: if 18,000 of those 20,000 input tokens are a stable prefix — system prompt, tool definitions, reference documents — the input cost per turn drops from $0.04 to about $0.0076.
Total per turn falls from 4.5 cents to about 1.3 cents, and the monthly bill from $1,350 to roughly $378 — a 72% reduction with no change to what the agent produces. The catch is that caching is a prefix match: any byte change anywhere in the prefix invalidates everything after it. A timestamp in the system prompt silently destroys the saving.
Batch processing halves it again for anything not latency-sensitive. Overnight classification, bulk extraction and backfills have no business being on the synchronous path.
Then the tradeoffs. Model tiering — routing the easy majority to a cheaper model and escalating the hard minority — and effort settings that let one model spend fewer tokens on routine work. Both trade some quality for cost, so measure on real traffic before rolling either out.
When does an agent stop paying for itself?
When the fully-loaded cost per task exceeds what the task is worth, which is a different calculation from cost per API call. Judge cost per completed task: a cheaper request that needs three retries and a human review is not cheaper.
The honest break-even test has three inputs. What did this work cost before, measured rather than estimated. What does the agent cost fully loaded, including review time and maintenance. And what share of volume does it actually handle end to end without escalating.
Production deployments report around 6.4 hours saved per worker per week, which is a real number and a biased one — it is drawn from the deployments that succeeded. The 88% that stalled are absent from every dataset of this kind.
The automations that clear the bar reliably share a shape: high volume, rule-heavy, low-judgement, with a human checkpoint that already existed. We wrote that test out in full in which AI automation applications actually reach production.
The short version
Token cost is the floor, not the bill. Model choice and prompt caching move it more than any amount of prompt tuning, and both are decisions you make once. Everything above the floor — retries, loop depth, review time, the data work underneath — is where agents actually become expensive.
Price the whole system before building, not the API call. If you want the fully-loaded version for a specific workflow, that is what the automation blueprint produces: a costed model of what it takes to run, not just what it takes to build.